LLM Knowledge Base
A prompt and methodology repository for extracting a three-layer knowledge model, scenarios, concepts and entities, from raw articles, then answering questions strictly from that modeled knowledge with sources cited.
LLM Knowledge Base is a prompt-and-reference repository, inspired by Andrej Karpathy's LLM-wiki practice, for building a personal or team knowledge base on top of raw source articles. It organizes knowledge into three layers: a scenario layer describing the problem being solved, a concept layer describing reusable units of thinking, modeled as an input-process-output description, a decomposition into sub-concepts, or both, and an entity layer describing concrete people, tools, frameworks or products.
Extraction runs in two passes: a fast first pass scans all source articles to build a deduplicated candidate list of scenarios, concepts and entities, and a second pass fills in definitions, relationships, sources and structure for each one, tracked in YAML files. A separate question-answering prompt then reads only the modeled scenarios, concepts and entities plus the raw source articles to build a structured, sourced answer, refusing to invent concepts absent from the model or use knowledge outside the raw articles. The repository is model, vector-database and framework agnostic, provides three offline HTML architecture diagrams, and is MIT licensed.